OpenAI has published the first benchmark results for Jalapeno, its first custom inference chip, and says it plans to begin deploying the system inside its own compute infrastructure by the end of 2026.
The company says Jalapeno delivered 1.5x to 1.9x more AI work per watt at peak throughput and 1.7x to 3.6x lower end-to-end latency than comparison systems across tests using GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T. For highly interactive workloads, OpenAI says performance was 2.1x to 4.1x higher.
Those are vendor-published results, and they should be read that way. The important news is still concrete: OpenAI now has measured first-party inference silicon claims, a public benchmark frame, and a near-term deployment target.
The benchmark is about useful work per watt
OpenAI says it tested Jalapeno on InferenceX, a SemiAnalysis benchmark that measures the full process of serving an AI request. The company argues that the relevant comparison is not only raw tokens per second per chip, but useful AI work completed per unit of power while meeting an interactive latency target.
That framing matters for agents. A multi-step agent compounds latency across planning, tool calls, code execution, retrieval, and response generation. A small delay in one request can become a visible delay when the agent performs dozens of sequential actions.
OpenAI says Jalapeno is rated at 700 watts, while measured sustained power stayed at or below 550 watts on the tested workloads. In the published appendix, the company reports a 1.03-second end-to-end latency on GPT-OSS 120B against 1.80 seconds for a GB200 comparison, and 1.65 seconds on DeepSeek R1 against 5.99 seconds for a GB300 comparison.
The Kimi K2.5 1T test is the most useful stress case in the post because it puts the chip against a large open-weight model outside OpenAI’s own family. OpenAI says Jalapeno delivered roughly 1.5x higher peak performance per watt and 3.4x lower end-to-end latency than the comparison system on that workload.
First-party silicon changes OpenAI’s supplier posture
Jalapeno does not mean OpenAI stops using NVIDIA or other accelerator partners. The company explicitly says it will continue to deploy accelerators from NVIDIA and other partners for training and inference.
The strategic shift is optionality. First-party inference silicon gives OpenAI a way to tune hardware, serving software, memory placement, networking, model architecture, and product demand as one system. That is different from buying general-purpose accelerator capacity and adapting serving software around it.
The company says Jalapeno was designed to handle both prefill and decode. Prefill is the stage where the system processes the prompt and accumulated context. Decode is the stage where it generates tokens. Agent workloads strain both phases because they often combine large context with repeated actions and user-visible response steps.
OpenAI also says AI helped develop and program the chip. The specific claim is bounded: for selected GPT-OSS attention and mixture-of-experts blocks, AI-generated implementations ran 1.5x to 1.8x faster than existing human-expert-written implementations. That is not a full-model speedup, but it points to a tighter loop between model development and systems optimization.
